Deep learning-based power transmission and transformation project standardized startup inspection method and system
By applying improved deep learning methods in the standardized start-up inspection of power transmission and transformation projects, including improved BERT model of syntax weights, white shark optimization algorithm and improved YOLOv5 algorithm, the problems of inaccurate and inefficient inspection results are solved, and more accurate and efficient inspection results are achieved.
Patent Information
- Application Number
- CN202411704273.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-04-15
AI Technical Summary
The prior art has problems of inaccurate inspection results and low efficiency in the standardized start-up inspection of power transmission and transformation projects, especially in text feature extraction and diversified inspection content processing.
Using a deep learning-based method, we collect and preprocess the construction approval procedures documents and on-site images, use the improved BERT model and White Shark optimization algorithm for text feature extraction and automatic inspection, and combine the improved YOLOv5 algorithm for image recognition and automatic inspection.
The accuracy and efficiency of standardized start-up inspections of power transmission and transformation projects have been improved, ensuring the reliability of inspection results and efficient processing of diversified inspection content.
Smart Images
Figure CN119831392A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a construction inspection method, which belongs to the technical field of power transmission and transformation engineering, and in particular to a standardized construction inspection method and system for power transmission and transformation engineering based on deep learning. Background Art
[0002] With the continuous advancement of industrialization and the continuous improvement of urbanization, the demand for electricity has shown a rapid growth trend. The safety, efficiency and stability of power supply have become key factors in ensuring the sustainable development of social economy. As an important part of the power system, the construction quality and efficiency of power transmission and transformation projects directly affect the reliability and stability of power supply. Standardized start-up inspections, as an important part of the construction of power transmission and transformation projects, can ensure that all preparations are in place in the early stages of the project, providing strong guarantees for the high-quality and efficient construction of the project. However, the contents that need to be inspected are complex and diverse, involving drawings, approval documents, on-site conditions and other aspects. This complexity often leads to low efficiency of inspection work, prone to omissions and errors, thus affecting the overall quality and progress of the project. Among the existing technical solutions, there is a power equipment defect text classification model based on BERT's dual-branch feature fusion, which identifies and classifies abnormal data in operation and maintenance data, or completes business process compliance inspection through improved BERT and lightweight CNN models; however, the above methods still have the problem of insufficient semantic information extraction in text feature extraction, which leads to deviations in inspection results, and relying solely on text file inspection cannot meet the needs of diversified inspection content in power transmission and transformation projects, and the inspection efficiency is low. Therefore, an accurate and efficient method is urgently needed to solve the above defects in the existing technology. Summary of the invention
[0003] The purpose of the present invention is to overcome the defects and problems of inaccurate inspection results and low efficiency in the prior art, and to provide an accurate and efficient standardized start-up inspection method and system for power transmission and transformation projects based on deep learning.
[0004] To achieve the above objectives, the technical solution of the present invention is: a standardized start-up inspection method for power transmission and transformation projects based on deep learning, comprising:
[0005] S1. Collect project construction approval documents and construction site images, and perform standardized preprocessing to obtain the text content of standardized start-up conditions for power transmission and transformation projects and the content of construction site images;
[0006] S2. Based on the text content of the standardized start-up conditions for power transmission and transformation projects, the BERT model is improved by introducing grammatical weights to calculate and obtain the text feature vector;
[0007] S3. Improve the White Shark optimization algorithm based on Sine-Tent-Cosine chaotic mapping and the position update strategy of the Firefly algorithm, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text;
[0008] S4, based on the improved YOLOv5 algorithm, identify the content of the construction site image and complete the automatic inspection of the construction site situation;
[0009] The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
[0010] The engineering construction approval procedure documents include: administrative approval procedures, company management requirements, internal approval procedures, and construction site conditions;
[0011] The administrative approval procedures include: project contract, construction permit, land use approval, and quality supervision registration;
[0012] The company management requirements include: signing contracts and responsibility commitment letters;
[0013] The internal approval procedures include: work commencement application form, process approval form;
[0014] The construction site conditions include: clarity of signs and plaques, availability of equipment and transportation, and access to construction manpower and machinery.
[0015] The step S2 specifically includes:
[0016] S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows:
[0017]
[0018] Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector;
[0019] S22, averaging the weight of each word to obtain the grammatical weight of each word; the expression of the grammatical weight is as follows:
[0020]
[0021] in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures;
[0022] S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector;
[0023]
[0024] Where: F is the text feature vector, is the weight of the self-attention mechanism.
[0025] In step S3, the improved White Shark optimization algorithm includes:
[0026] S31. Optimize the white shark algorithm based on Sine-Tent-Cosine chaotic mapping to generate an initial population of white sharks; the expression of the initial population is as follows:
[0027]
[0028] Where: x(i) is the position of the white shark population in the i-th iteration, x(i+1) is the position of the white shark population in the i+1-th iteration;
[0029] S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows:
[0030]
[0031] in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark;
[0032]
[0033] in: is the position of the kth i-th white shark, α2 is the control parameter;
[0034] S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows:
[0035]
[0036] Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
[0037] In step S4, the improved YOLOv5 algorithm specifically includes:
[0038] S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows:
[0039]
[0040] Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w , C h The diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively;
[0041] S42. Introduce the balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows:
[0042] L EIOU ′=IOU α L EIOU ;
[0043] Among them: IOU α is a balancing term, and α is a hyperparameter that controls the loss value.
[0044] In step S4, the CBAM attention mechanism in the YOLOv5 algorithm is improved, which means:
[0045] S43, by accepting two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOLOv5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention M s1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2 F in1 、F in2After adjustment, they are recombined to output the final feature map; the expression of the recombined after adjustment is as follows:
[0046]
[0047] F out =C out +S out ;
[0048] Where: C out is the feature map output after channel attention adjustment, S out is the feature map output after spatial attention adjustment, is element-wise multiplication.
[0049] A standardized start-up inspection system for power transmission and transformation projects based on deep learning, the system comprising:
[0050] The document image module is used to collect project construction approval documents and construction site images, and perform standardized preprocessing to obtain the text content of the standardized start-up conditions of the power transmission and transformation project and the construction site image content;
[0051] The text feature module is used to improve the BERT model by introducing grammatical weights based on the text content of the standardized start-up conditions of power transmission and transformation projects, and calculate the text feature vector;
[0052] The text inspection and classification module is used to improve the White Shark optimization algorithm based on the Sine-Tent-Cosine chaotic mapping and the firefly algorithm position update strategy, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text;
[0053] The construction site inspection module is used to identify the content of construction site images based on the improved YOLOv5 algorithm and complete automatic inspection of the construction site conditions;
[0054] The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
[0055] The text feature module is used to improve the BERT model according to the following steps:
[0056] S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows:
[0057]
[0058] Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector;
[0059] S22, averaging the weight of each word to obtain the grammatical weight of each word; the expression of the grammatical weight is as follows:
[0060]
[0061] in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures;
[0062] S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector;
[0063]
[0064] Where: F is the text feature vector, is the weight of the self-attention mechanism.
[0065] The text inspection classification module is used to improve the White Shark optimization algorithm according to the following steps:
[0066] S31. Optimize the white shark algorithm based on Sine-Tent-Cosine chaotic mapping to generate an initial population of white sharks; the expression of the initial population is as follows:
[0067]
[0068] Where: x(i) is the position of the white shark population in the i-th iteration, x(i+1) is the position of the white shark population in the i+1-th iteration;
[0069] S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows:
[0070]
[0071] in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark;
[0072]
[0073] in: is the position of the kth i-th white shark, α2 is the control parameter;
[0074] S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows:
[0075]
[0076] Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
[0077] The construction site inspection module is used to improve the YOLOv5 algorithm according to the following steps:
[0078] S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows:
[0079]
[0080] Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w , C h The diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively;
[0081] S42. Introduce the balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows:
[0082] L EIOU ′=IOU α L EIOU ;
[0083] Among them: IOU α is a balance term, and α is a hyperparameter that controls the loss value;
[0084] S43, by accepting two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOLOv5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention Ms1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2 F in1 、F in2 After adjustment, they are recombined to output the final feature map; the expression of the recombined after adjustment is as follows:
[0085]
[0086] F out =C out +S out ;
[0087] Where: C out is the feature map output after channel attention adjustment, S out is the feature map output after spatial attention adjustment, is element-wise multiplication.
[0088] Compared with the prior art, the present invention has the following beneficial effects:
[0089] The invention discloses a method and system for standardized start-up inspection of power transmission and transformation projects based on deep learning. The method first collects project construction approval procedure documents and construction site images, performs standardized preprocessing, obtains text content and site images, then introduces grammatical weights to improve the BERT model based on the text content, calculates and obtains text feature vectors, then improves the White Shark optimization algorithm based on Sine-Tent-Cosine chaotic mapping and firefly algorithm position update strategy, takes the text feature vector as input, outputs automatic inspection and classification results of the text, and finally recognizes the content of the construction site image based on the improved YOLOv5 algorithm to complete the automatic inspection of the construction site situation. In the application of the design, the feature vector representation of the text is enhanced by optimizing the self-attention mechanism of BERT, and the convergence speed of BiLSTM is enhanced by using the improved White Shark optimization algorithm, while the recognition ability of the YOLOv5 algorithm is improved, and the image is combined with the text to complete the standardized inspection of the construction site, which not only ensures the accuracy of the inspection result, but also improves the efficiency of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0090] Figure 1 It is a flow chart of the method steps of the present invention.
[0091] Figure 2 It is a schematic diagram of the system structure of the present invention.
[0092] Figure 3 It is a schematic diagram of the device structure of the present invention.
[0093] In the figure: document image module 1, text feature module 2, text inspection and classification module 3, construction site inspection module 4, processor 5, memory 6, computer program code 61. DETAILED DESCRIPTION
[0094] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0095] Embodiment 1:
[0096] See also Figure 1 , a standardized start-up inspection method for power transmission and transformation projects based on deep learning, including:
[0097] S1. Collect project construction approval documents and construction site images, and perform standardized preprocessing to obtain the text content of standardized start-up conditions for power transmission and transformation projects and the content of construction site images;
[0098] Furthermore, by querying and sorting out the rules and regulations and management requirements for standardized commencement of construction, and based on actual work needs, this plan comprehensively selects four aspects for inspection: administrative approval procedures before commencement of construction, company management requirements that should be implemented, internal approval procedures that should be performed, and on-site conditions that should be met at the construction site.
[0099] The administrative approval procedures include: project contract, construction permit, land use approval, and quality supervision registration;
[0100] The company management requirements include: signing contracts and responsibility commitment letters;
[0101] The internal approval procedures include: work commencement application form, process approval form;
[0102] The construction site conditions include: clarity of signs and plaques, availability of equipment and transportation, and access to construction manpower and machinery.
[0103] The standardized preprocessing refers to preprocessing the collected data to construct a standardized data format. The specific steps include:
[0104] The text content is cleaned and filtered based on the NLTK module to remove redundant spaces and line breaks in the document, making subsequent text analysis more efficient and accurate. The Imgaug enhancement library is used to flip and rotate the construction site images, so that the model can learn features from multiple perspectives.
[0105] S2. Based on the text content of the standardized start-up conditions for power transmission and transformation projects, the BERT model is improved by introducing grammatical weights to calculate and obtain the text feature vector;
[0106] Furthermore, the step S2 specifically includes:
[0107] S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism, and converts the words into vectors by obtaining the weight relationship between the words to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows:
[0108]
[0109] Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector;
[0110] Since each sentence contains words of different structures, the representation ability of words will directly affect the representation effect of sentences. Therefore, in order to improve the representation effect of generated sentences, this solution introduces the structure information of sentences on the basis of the self-attention mechanism, taking into account the weight of each word in all grammatical components in the sentence. By combining the grammatical component weights of words with the self-attention mechanism, the model can understand and represent the complex structure of sentences more accurately.
[0111] S22. For a sentence containing M structures, this solution averages the weight of each word in the sentence to obtain the grammatical weight of the word, which can be specifically expressed as follows:
[0112]
[0113] in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures;
[0114] S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector;
[0115]
[0116] Where: F is the text feature vector, is the weight of the self-attention mechanism.
[0117] S3. Improve the White Shark optimization algorithm based on Sine-Tent-Cosine chaotic mapping and the position update strategy of the Firefly algorithm, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text;
[0118] The result of the automatic inspection is whether the text file is compliant, and the compliant and non-compliant text files are automatically classified to obtain the classification result.
[0119] Furthermore, based on the text feature vector extracted by BERT as input, this solution completes automatic inspection based on the improved BiLSTM; BiLSTM combines forward and backward long short-term memory networks to capture the features of the learned text vector, and finally realizes automatic inspection and classification of documents through the Softmax function in the output layer.
[0120] Furthermore, the improved White Shark optimization algorithm includes:
[0121] S31. Aiming at the problem of uneven distribution of initial population in WSO algorithm, which leads to insufficient search capability, this solution optimizes the White Shark algorithm based on Sine-Tent-Cosine chaotic mapping, combining the different distribution characteristics of Sine, Tent and Cosine mapping to generate a more uniform and diversified initial population; the expression of the initial population is as follows:
[0122]
[0123] Where: x(i) is the position of the white shark population in the i-th iteration, x(i+1) is the position of the white shark population in the i+1-th iteration;
[0124] S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows:
[0125]
[0126] in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark;
[0127]
[0128] in: is the position of the kth i-th white shark, α2 is the control parameter;
[0129] The WSO algorithm is prone to fall into the problem of local optimality in the later iteration process. Therefore, this scheme improves it based on the position update strategy of the firefly algorithm (FA), which improves the accuracy and ability to escape the local optimality to a certain extent. The proposed method introduces the attraction formula in FA into the WSO algorithm to further guide the white shark to find the best individual, and adds a disturbance term at the end to increase the randomness of the algorithm.
[0130] S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows:
[0131]
[0132] Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
[0133] S4, based on the improved YOLOv5 algorithm, identify the content of the construction site image and complete the automatic inspection of the construction site situation;
[0134] In this embodiment, the construction site image content based on the YOLOv5 algorithm is to obtain specific objects related to standardized inspection, such as equipment, tool equipment, vehicles, construction machinery entry, etc. in the construction site conditions, so as to check whether the construction site meets the standards for standardized start-up. The objects of inspection are objects related to standardized inspection identified by the YOLOv5 model.
[0135] Based on the objects detected by the YOLOv5 model, the system will compare these objects with the standards of the construction site to complete the automatic inspection. The output of the YOLOv5 model is the detection result with the marked box and category. During the inspection, for example, for the entry of mechanical equipment to the construction site, the system checks whether there are detection boxes of construction machinery categories such as bulldozers and excavators in the content output by the YOLOv5 model in the corresponding construction area. If there are, it is compliant, otherwise it is non-compliant; the result of the inspection is the compliance or non-compliance of the corresponding inspection object.
[0136] The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
[0137] The CIOU loss function introduces an additional aspect ratio difference as a penalty term based on the calculation of the center point offset between the true box and the predicted box. However, this penalty term only considers the overall difference in aspect ratio, but does not consider the difference between the width and height of the predicted box and the true box. Therefore, this solution introduces the EIOU loss function, which directly uses the prediction results of width and height as penalty terms to improve the model accuracy.
[0138] Furthermore, the improved YOLOv5 algorithm specifically includes:
[0139] S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows:
[0140]
[0141] Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w , C h The diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively;
[0142] S42. In order to avoid the problem of loss value fluctuation caused by uneven sample quality during training, this solution introduces a balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows:
[0143] L EIOU ′=IOU α L EIOU ;
[0144] Among them: IOU α is a balancing term, and α is a hyperparameter that controls the loss value.
[0145] By introducing the balancing term IOU α , which can suppress the contribution of low-quality samples in the loss value, avoid the problem that low-quality samples dominate the loss and cause the loss function to fail to converge quickly, and make the model pay more attention to high-quality samples, thereby further increasing the convergence speed of the model.
[0146] The improvement of the CBAM attention mechanism in the YOLOv5 algorithm refers to:
[0147] S43, CBAM attention mechanism highlights the important structures and details in the image through serially connected channels and spatial attention. This scheme accepts two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOLOv5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention M s1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2F in1 、F in2 After adjustment, they are recombined to obtain a more accurate feature representation and output the final feature map; the expression after adjustment and recombination is as follows:
[0148]
[0149] F out =C out +S out ;
[0150] Where: C out is the feature map output after channel attention adjustment, S out is the feature map output after spatial attention adjustment, is element-wise multiplication.
[0151] This scheme calculates attention through feature maps after multi-scale fusion, which can make the information of feature maps of different scales complement each other, reduce the information loss that may occur in the processing process, and improve the performance of the model in capturing detailed information; at the same time, the parallel processing method allows channel and spatial attention to be performed simultaneously, avoiding the sequential transmission in serial processing, and can retain more original feature information, which helps to capture more details and reduce the overall calculation time.
[0152] Embodiment 2:
[0153] See also Figure 2 , a standardized start-up inspection system for power transmission and transformation projects based on deep learning, the system comprising:
[0154] Document image module 1 is used to collect project construction approval documents and construction site images, and perform standardized preprocessing to obtain the text content of the standardized start-up conditions of the power transmission and transformation project and the construction site image content;
[0155] Furthermore, the engineering construction approval procedure documents collected by the document image module 1 include: administrative approval procedures, company management requirements, internal approval procedures, and construction site conditions;
[0156] The administrative approval procedures include: project contract, construction permit, land use approval, and quality supervision registration;
[0157] The company management requirements include: signing contracts and responsibility commitment letters;
[0158] The internal approval procedures include: work commencement application form, process approval form;
[0159] The construction site conditions include: clarity of signs and plaques, availability of equipment and transportation, and access to construction manpower and machinery.
[0160] Text feature module 2 is used to improve the BERT model by introducing grammatical weights based on the text content of the standardized start-up conditions of power transmission and transformation projects, and calculate the text feature vector;
[0161] Furthermore, the text feature module 2 is used to improve the BERT model according to the following steps:
[0162] S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows:
[0163]
[0164] Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector;
[0165] S22, averaging the weight of each word to obtain the grammatical weight of each word; the expression of the grammatical weight is as follows:
[0166]
[0167] in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures;
[0168] S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector;
[0169]
[0170] Where: F is the text feature vector, is the weight of the self-attention mechanism.
[0171] Text inspection and classification module 3 is used to improve the White Shark optimization algorithm based on the Sine-Tent-Cosine chaotic mapping and the firefly algorithm position update strategy, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text;
[0172] Furthermore, the text inspection classification module 3 is used to improve the White Shark optimization algorithm according to the following steps:
[0173] S31. Optimize the white shark algorithm based on Sine-Tent-Cosine chaotic mapping to generate an initial population of white sharks; the expression of the initial population is as follows:
[0174]
[0175] Where: x(i) is the position of the white shark population in the i-th iteration, x(i+1) is the position of the white shark population in the i+1-th iteration;
[0176] S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows:
[0177]
[0178] in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark;
[0179]
[0180] in: is the position of the kth i-th white shark, α2 is the control parameter;
[0181] S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows:
[0182]
[0183] Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
[0184] The construction site inspection module 4 is used to identify the content of the construction site image based on the improved YOLOv5 algorithm and complete the automatic inspection of the construction site conditions;
[0185] The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
[0186] Furthermore, the construction site inspection module 4 is used to improve the YOLOv5 algorithm according to the following steps:
[0187] S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows:
[0188]
[0189] Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w , C h The diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively;
[0190] S42. Introduce the balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows:
[0191] L EIOU ′=IOU α L EIOU ;
[0192] Among them: IOU α is a balance term, and α is a hyperparameter that controls the loss value;
[0193] S43, by accepting two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOLOv5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention M s1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2 F in1 、F in2 After adjustment, they are recombined to output the final feature map; the expression of the recombined after adjustment is as follows:
[0194]
[0195] F out =C out +S out ;
[0196] Where: C out is the feature map output after channel attention adjustment, S outis the feature map output after spatial attention adjustment, is element-wise multiplication.
[0197] Embodiment 3:
[0198] See also Figure 3 In this embodiment, a standardized start-up inspection device for power transmission and transformation projects based on deep learning is also included, and the device includes a processor 5 and a memory 6;
[0199] The memory 6 is used to store the computer program code 61 and transmit the computer program code 61 to the processor 5;
[0200] The processor 5 is used to execute the standardized commencement inspection method for power transmission and transformation projects based on deep learning described in Example 1 according to the instructions in the computer program code 61.
[0201] This embodiment also includes a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed on a computer, the standardized commencement inspection method for power transmission and transformation projects based on deep learning described in Example 1 is implemented.
[0202] Generally speaking, the computer instructions for implementing the method of the present invention may be carried in any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media may include any computer-readable media, except for the signal itself that is temporarily propagating.
[0203] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EKROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, device, or device.
[0204] Computer program code for performing the operation of the present invention can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, SMalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages, in particular, Python suitable for neural network computing and platform frameworks based on TensorFlow, PyTorch, etc. can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer or to an external computer (for example, using an Internet service provider to connect via the Internet) through any type of network, including a local area network (LAN) or a wide area network (WAN).
[0205] The above-mentioned devices and non-temporary computer-readable storage media can be found in the specific description of a standardized commencement inspection method for power transmission and transformation projects based on deep learning and its beneficial effects, which will not be repeated here.
[0206] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.
Claims
1. A standardized start-up inspection method for power transmission and transformation projects based on deep learning, characterized in that: include: S1. Collect project construction approval documents and construction site images, and perform standardized preprocessing to obtain the text content of standardized start-up conditions for power transmission and transformation projects and the content of construction site images; S2. Based on the text content of the standardized start-up conditions for power transmission and transformation projects, the BERT model is improved by introducing grammatical weights to calculate and obtain the text feature vector; S3. Improve the White Shark optimization algorithm based on Sine-Tent-Cosine chaotic mapping and the position update strategy of the Firefly algorithm, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text; S4, based on the improved YOLOv5 algorithm, identify the content of the construction site image and complete the automatic inspection of the construction site situation; The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
2. The method for standardized start-up inspection of power transmission and transformation projects based on deep learning according to claim 1 is characterized in that: The engineering construction approval procedure documents include: administrative approval procedures, company management requirements, internal approval procedures, and construction site conditions; The administrative approval procedures include: project contract, construction permit, land use approval, and quality supervision registration; The company management requirements include: signing contracts and responsibility commitment letters; The internal approval procedures include: work commencement application form, process approval form; The construction site conditions include: clarity of signs and plaques, availability of equipment and transportation, and access to construction manpower and machinery.
3. The method for standardized start-up inspection of power transmission and transformation projects based on deep learning according to claim 1 is characterized in that: The step S2 specifically includes: S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows: Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector; S22, averaging the weight of each word to obtain the grammatical weight of each word; the expression of the grammatical weight is as follows: in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures; S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector; Where: F is the text feature vector, is the weight of the self-attention mechanism.
4. The method for standardized start-up inspection of power transmission and transformation projects based on deep learning according to claim 1 is characterized in that: In step S3, the improved White Shark optimization algorithm includes: S31. Optimize the white shark algorithm based on Sine-Tent-Cosine chaotic mapping to generate an initial population of white sharks; the expression of the initial population is as follows: Where: x(i) is the position of the white shark population in the i-th iteration, x(i+1) is the position of the white shark population in the i+1-th iteration; S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows: in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark; in: is the position of the kth i-th white shark, α2 is the control parameter; S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows: Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
5. The method for standardized start-up inspection of power transmission and transformation projects based on deep learning according to claim 1 is characterized in that: In step S4, the improved YOLOv5 algorithm specifically includes: S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows: Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w , C h The diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively; S42. Introduce the balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows: L EIOU ′=iOU α L EIOU ; Among them: IOU α is a balancing term, and α is a hyperparameter that controls the loss value.
6. The method for standardized start-up inspection of power transmission and transformation projects based on deep learning according to claim 1 is characterized in that: In step S4, the CBAM attention mechanism in the YOLOv5 algorithm is improved, which means: S43, by accepting two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOLOv5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention M s1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2 F in1 、F in2 After adjustment, they are recombined to output the final feature map; the expression of the recombined after adjustment is as follows: F out =C out +S out ; Where: C out is the feature map output after channel attention adjustment, S out is the feature map output after spatial attention adjustment, is element-wise multiplication.
7. A standardized start-up inspection system for power transmission and transformation projects based on deep learning, characterized in that: The system comprises: The document image module (1) is used to collect the project construction approval procedure documents and the construction site images, and perform standardized pre-processing to obtain the text content of the standardized start-up conditions of the power transmission and transformation project and the construction site image content; A text feature module (2) is used to improve the BERT model by introducing grammatical weights based on the text content of the standardized start-up conditions for power transmission and transformation projects, and calculate and obtain a text feature vector; The text inspection and classification module (3) is used to improve the White Shark optimization algorithm based on the Sine-Tent-Cosine chaotic mapping and the firefly algorithm position update strategy, and use the text feature vector as the input of the BiLSTM network to output the automatic inspection and classification results of the text; A construction site inspection module (4), used to identify the content of the construction site image based on the improved YOLOv5 algorithm, and complete automatic inspection of the construction site conditions; The improved YOLOv5 algorithm refers to: introducing a balance term to improve the EIOU loss function, introducing the improved EIOU loss function into the YOLOv5 algorithm, and improving the CBAM attention mechanism in the YOLOv5 algorithm.
8. The deep learning-based standardized start-up inspection system for power transmission and transformation projects according to claim 7 is characterized by: The text feature module (2) is used to improve the BERT model according to the following steps: S21. The BERT model splits the sentences in the text content of the standardized start-up conditions for power transmission and transformation projects into words, and calculates the relevance of each word with other words in the context based on the self-attention mechanism to obtain the self-attention mechanism weight; the calculation formula of the self-attention mechanism is as follows: Where: Q, K, V are the vectors of query, key, and value respectively, d k is the dimension of the input word vector; S22, averaging the weight of each word to obtain the grammatical weight of each word; the expression of the grammatical weight is as follows: in: is the grammatical weight of the i-th word, is the weight of the i-th word on the k-th grammatical component, and M is a sentence with M structures; S23, adding the grammatical weight to the self-attention mechanism weight to obtain the text feature vector; Where: F is the text feature vector, is the weight of the self-attention mechanism.
9. The deep learning-based standardized start-up inspection system for power transmission and transformation projects according to claim 7 is characterized by: The text inspection classification module (3) is used to improve the White Shark optimization algorithm according to the following steps: S31. Optimize the white shark algorithm based on Sine-Tent-Cosine chaotic mapping to generate an initial population of white sharks; the expression of the initial population is as follows: Where: x( i ) is the position of the white shark population in the i-th iteration, and x(i+1) is the position of the white shark population in the i+1-th iteration; S32. Using the clustering behavior of white sharks, simulate the group moving toward the white shark closest to the prey and update the position of the white sharks in the initial population; the expression for position update is as follows: in: is the position of the i-th white shark in the k+1th generation, is the optimal position of the white shark up to the kth iteration, r1, r2, r3 are random numbers in [0, 1], sgn(r2-0.5) is the search direction of the white shark, is the distance between the white shark and its prey, S s are the olfactory and visual intensity parameters of the white shark; in: is the position of the kth i-th white shark, α2 is the control parameter; S33. The White Shark algorithm is optimized based on the location update strategy of the Firefly algorithm. The optimized location update expression is as follows: Among them: β0 is the initial attraction, γ is the attraction coefficient, r is the distance between the white shark and the optimal white shark, α is the step size, and ε is a random number between [-0.5, 0.5].
10. The deep learning-based standardized start-up inspection system for power transmission and transformation projects according to claim 7 is characterized by: The construction site inspection module (4) is used to improve the YOLOv5 algorithm according to the following steps: S41. Replace the CIOU loss function in the YOLOv5 algorithm with the EIOU loss function; the expression of the EIOU loss function is as follows: Where: L IoU is the intersection-over-union loss, L dis is the center point offset loss, L asp is the width-height loss, ρ 2 is the Euclidean distance, b, w, h are the center coordinates, width and height of the prediction box, c, C w、 Ch is the diagonal distance, width and height of the minimum bounding rectangle of the predicted box and the real box respectively; S42. Introduce the balance term to improve the EIOU loss function; the improved EIOU loss function expression is as follows: L EIOU ′=IOU α L EIOU ; Among them: IOU α is a balance term, and α is a hyperparameter that controls the loss value; S43, by accepting two feature maps of different scales F in1 、F in2 After fusion, it is input into the YOL0v5 algorithm to generate the channel attention M corresponding to the two different scale feature maps of the input c1 、M c2 and spatial attention M s1 、M s2 , and through the channel attention M c1 、M c2 and spatial attention M s1 、M s2 F in1 、F in2 After adjustment, they are recombined to output the final feature map; the expression of the recombined after adjustment is as follows: F out =C out +S out ; Where: C out , is the feature map output after channel attention adjustment, S out is the feature map output after spatial attention adjustment, is element-wise multiplication.